EAFNet: Feature Enhancement and Self-Adaptive Guided Feature Fusion for Object Detection in Haze Conditions

Recently, there is a significant success in the current object detection task. However, existing research has demonstrated that detectors trained on regular images experience a significant decrease when facing to wild scences such as haze scence. Although enhancing detection performance is possible through multi-stage adaptive parameter-guided image enhancement, this method fails to consistently enhance downstream object detection performance. To address this issue, we propose an end-to-end object detection framework that utilizes weakly supervised parameter-guided feature fusion to improve detection performance. Specifically, we introduce the Self-Adaptive Feature Fusion Module (SAFFM) that dynamically guides feature fusion in the detection network through the haze image information encoded by the Prior Knowledge (PK) components. By the joint learning of the PK components and the detection network in an end-to-end manner, we ensure that the PK components weakly supervise the feature fusion process to enhance detection performance. Additionally, we propose a Downsampling Feature Enhancement Module (DFEM) that utilizes multibranch down-sampling operators and reparameterization to enhance the feature extraction capability of the network, allowing it to capture more latent information. Comprehensive experiments on synthetic datasets and real datasets, demonstrated that the detection performance of the proposed network significantly outperforms other state-of-the-art detection methods.